Using prediction results, selecting images that are most informative. These images tend to have inconsistent confidence score among different classes.
Four methods are supported in this program: (the first two methods are visually better)
- Max Entropy (single model)
- Sum Entropy (single model)
- Mutual Information (multi models)
- KL divergence on score distribution (multi models)
For detailed description of the method, please read this documentation.
Command line arguments include:
--method, -m: Active learning method. Choose one: sum_entropy, max_entropy, mutual_information or KL_score--pred_path, -p: Paths to prediction results. You can pass in predictions from one or multiple models, separating by commas. Eg: ./pred1.json,./pred2.json--gt_path, -g: Path to ground truth annotations in coco format--image_root, -i: Directory containing images--visualize, -v: Flag to visualize selected image. Default = True--num_query, -n: Number of selected images
Example python command to run the program with max entropy
python main.py -m max_entropy -p predictions.json -n 20 -i ./images -g groundtruths.json